Submarine landslides are a major geohazard, often preconditioned by micro-structural heterogeneity within sensitive weak layers. While X-ray micro-computed tomography (μCT) offers high-resolution insights into these sediments, quantifying phase distributions remains a bottleneck due to the complexity of manual segmentation. This study establishes an end-to-end deep learning workflow for the automated three-phase segmentation and 3D quantitative analysis of submarine sediments. Specifically, we compared two architectures, U-Net and DeepLabV3+, on high-resolution μCT scans of slope sediments from Finneidfjord (Norway) and the AFEN slide (UK). Both models achieved high segmentation accuracy (Dice > 0.84). DeepLabV3+ demonstrated superior computational efficiency and generalization for gas phases, while U-Net proved more robust in resolving subtle liquid–solid interfaces crucial for pore-scale mechanics. To evaluate model robustness and generalization, both approaches were tested across contrasting sedimentary endmembers, including homogeneous background sediments and micro-structurally heterogeneous weak-layer sequences relevant for slope failure initiation. The automated 3D reconstruction revealed a specific "coarse-grain framework with fine-grain infill" texture in the Finneidfjord weak layer, suggesting that obstructed drainage pathways may facilitate excess pore pressure accumulation. Ultimately, this study establishes an AI-driven workflow for high-throughput quantitative μCT analysis, enabling the identification of micro-mechanical precursors associated with slope failure.
When large language models are applied to professional geological fields such as mineralogy, they frequently suffer from knowledge obsolescence, factual hallucinations, and invalid complex logical reasoning. Existing retrieval-augmented generation (RAG) methods mostly adopt a static pipeline architecture, which easily causes logical breaks and semantic fragmentation when processing complex geological queries, and can hardly realize collaborative utilization of structured knowledge and unstructured text. To address the above shortcomings, this paper proposes a Multi-agent Mineral RAG (MaM-RAG) approach for mineralogy-oriented collaborative retrieval-augmented generation. This method constructs a collaborative closed loop of intent recognition-multi-source execution-decision fusion. An intent recognition agent performs non-destructive decoupling of complex queries while completely preserving the internal logical structure. Multi-source retrieval agents acquire knowledge graphs and text corpora in parallel to break information isolation among heterogeneous data. A decision fusion agent conducts semantic consistency verification on multi-source evidence and triggers reflective retrieval in case of evidence inconsistency, completing self-verification and self-correction of the reasoning process. Experimental results show that the proposed method achieves significantly higher accuracy than traditional RAG and general large language models on mineralogy professional question answering tasks. It can effectively suppress factual hallucinations and improve answer accuracy and robustness, providing a feasible solution for professional intelligent question answering in the geological domain.
Intelligent identification and characterization of sedimentary microfacies is the core content of oil and gas reservoir geological research, which has important guiding significance for the analysis of reservoir space distribution law and the optimization of oil and gas reservoir development scheme. Traditional interpretation methods rely on manual judgment of logging response parameters, limited by the difference in subjective experience of interpreters, and machine learning as a new method is gradually replacing some manual judgments. Since machine learning does not accurately determine all deposited sedimentary microfacies, it is necessary to adequately compare multiple models. In this study, taking the Cretaceous Yaojia Formation (K2y1) member of the Putaohua oil layer in the satellite oilfield of Songliao Basin as the research object, this study builds on existing research and proposes a deep learning sedimentary microfacies intelligent identification method based on the time series image features of logging curves, and constructs three typical models for comparison, namely CNN, ResNet and Transformer. By converting five logging curves, natural gamma (GR), natural potential (SP), acoustic time difference (AC), deep lateral resistivity (LLD) and shallow lateral resistivity (LLS), into temporal images, the morphological parameters reflecting sedimentary dynamics are extracted, and the correspondence between logging response characteristics and sedimentary microfacies is established. The research results show that CNN, ResNet, and Transformer have accuracy rates of 83%, 85%, and 90% respectively on the validation set. Through the comprehensive evaluation of multiple indicators, the applicable conditions and optimization directions of different models in the identification of sedimentary microfacies in heterogeneous reservoirs are clarified. This method system provides an improved approach for interpreting complex reservoir sedimentary microfacies, and provides a new technical approach for multi-model comparison and optimization for intelligent exploration of hidden oil and gas reservoirs.
Conventional numerical approaches to forward modeling in 3D airborne transient electromagnetic (ATEM) surveys are constrained by limited computational efficiency and prohibitively high costs, which severely limit the practical application of 3D inversion in exploration scenarios. Existing deep learning-based fast forward modeling approaches are largely restricted to low-dimensional problems and are incapable of handling full 3D forward simulations. To overcome this limitation, we propose a deep learning-based fast forward modeling framework that enables efficient and accurate 3D forward modeling. The model adopts a Transformer architecture as its backbone network. To reduce computational complexity, an octree embedding mechanism is introduced to adaptively discretize the resistivity model, fine-grained embeddings are employed in regions with strong resistivity variations, while coarse-grained representations are used in relatively homogeneous areas. This strategy significantly reduces computational redundancy and sequence length while preserving modeling accuracy. To account for the influence of transceiver altitude variations on forward responses, a transceiver altitude-gating mechanism is introduced, allowing the network to accurately perceive and respond to response variations under different transceiver altitudes. Comprehensive experimental results indicate that the proposed model enables fast, high-precision 3D forward modeling, exhibits strong sensitivity to transceiver altitude variations, and maintains robust adaptability when applied to realistic models.
To overcome the low efficiency of performing frequency-domain acoustic wavefield simulation, this paper proposes a kind of deep learning (DL) preconditioned methods based on sufficiently training deep neural networks to construct efficient preconditioners, which can accelerate BiCGSTAB iterative method for solving the linear system formed after discretizing the wave equation. The main novelty of these solvers is reflected in the combination of networks’ powerful representation capability with classical preconditioning to deal with the adverse characteristics of the impedance matrix. Compared with traditional approaches, the preconditioning process of DL preconditioned iteration is conducted by network prediction, which can converge to true solution with fewer iterative steps and less cost. Numerical examples of several typical media are implemented to examine the efficacy of the proposed methods. As a result, it demonstrates their superiority in accelerating wavefield simulation and is beneficial to enhance the feasibility and practicality of frequency-domain modeling in seismic exploration.
Hyperspectral image classification (HSI) is extensively utilized to analyze remotely sensed images for different real-world applications. Recently, convolutional neural network(CNN) have been applied successfully for HSI classification. Ensemble learning methods such as adaptive boosting (AdaBoost), random forest(RF), and extreme gradient boosting (XGBoost) are comparatively simpler techniques that also achieve a significant level of performance for HSI classification. This article proposes a hybrid classifier for HSI integrating the merits of two prominent classifiers: CNN and ensemble learning method. Both of them have evidence of efficient recognition capability of finding patterns from data. The ensemble model performs recognition using the salient features extracted by CNN. A modified version of AlexNet compatible with HSI datacube has been adopted as the CNN model. Three ensemble methods have been applied for recognition: AdaBoost, RF, and XGBoost. Two base learners have been used with AdaBoost: decision tree and Support Vector Machine(SVM). Experiments have been performed on three benchmark datasets: Indian Pines(IN), University of Pavia(UP) and Kennedy Space Centre(KSC), which mostly include land cover of agriculture, forest, soil, rural, and urban area. Experimental results establish the superiority of CNN with Adaboost hybrid method where SVM is used as the base learner. It delivers maximum overall accuracy (OA) of 91.48%, 99.18%, and 88.10% and kappa score of 90.25%, 98.91%, and 86.71%, for IN, UP and KSC dataset, respectively.
Elastic full-waveform inversion (EFWI) constitutes a vital tool for high-resolution subsurface imaging. However, its application in shallow-subsurface exploration is hindered by strong nonlinearity, extreme sensitivity to the initial model, and the low sensitivity of specific parameters, particularly density. To address these challenges, we introduce a Siamese CNN-based EFWI framework (SCFWI) for multi-parameter inversion. This framework embeds a weight-sharing Siamese network into the physical inversion loop, adaptively extracting multi-scale common features from both observed and synthetic seismic data. We construct a novel feature-data collaborative objective function that imposes dual constraints: minimizing feature discrepancies to recover global structures and constrain low-sensitivity parameters, while simultaneously reducing waveform residuals to preserve local high-resolution details. Numerical experiments demonstrate that SCFWI significantly mitigates reliance on accurate initial models and enhances noise robustness. Notably, the method achieves a 56.6% reduction in the density root-mean-square error (RMSE) for the SEAM model compared to conventional EFWI, highlighting its capability to mitigate parameter crosstalk. This study presents a promising physics-guided deep learning framework for multi-parameter EFWI, offering valuable insights for broader intelligent geophysical inversion applications.
This work presents a comparative study of twelve reinforcement learning (RL) algorithms for training-free neural architecture search (NAS) of fully connected neural networks (FCNNs) with skip connections. Neural architecture search is formulated as a multi-armed bandit problem represented as a Markov decision process, where states encode FCNN configurations (hidden layer widths, activation functions, skip connection topology) and actions comprise discrete architectural modifications (layer insertion/deletion, neuron adjustment, activation function change, skip connection establishment). The reward signal is the Negative Condition Number (NCN) metric derived from Neural Tangent Kernel theory, enabling training-free architecture evaluation. The study evaluates twelve RL agents: on-policy stochastic methods (Policy Gradient, Actor-Critic, A2C, A3C, PPO), on-policy trust-region methods (TRPO), off-policy deterministic methods (DDPG, TD3, SAC), and evolution strategy variants (DDPG-ES, TD3-ES, SAC-ES). Empirical validation uses a dataset of pyrolysis data from the HAWK experiment comprising Total Organic Carbon (TOC), Thermal Maturity (Tmax), Hydrogen Index (HI), and Oxygen Index (OI) from 114 cuttings across eight wells in a Persian Gulf source rock, matched with wireline log measurements: corrected gamma ray (CGR), neutron porosity (PHIN), bulk density (RHOB), compressional wave slowness (DT), and true formation resistivity (Rt). Results show on-policy methods achieved stable convergence with consistent reward improvement and decreasing exploration variance. PPO and A3C agents achieved better final reward after training. Off-policy methods (DDPG, TD3, SAC) and evolutionary strategy variants failed to converge, displaying oscillatory behavior or early stagnation in the high-dimensional action space. The PPO-discovered architecture (3724 parameters) exhibits lower cross-validation variance than the A3C-discovered architecture (83,944 parameters) with comparable bias after full training using K-fold cross validation. These findings establish on-policy RL methods as effective for FCNN architecture optimization in high-dimensional action spaces.
Reliable estimation for the horizontal seismic site amplification factors (HSAF) is an essential step in the deterministic seismic hazard analyses. The current study introduces newly hybrid machine learning (ML) framework to estimate the HSAF using solely the horizontal-to-vertical spectral ratios of earthquakes (EHVSR) and microtremors (MHVSR). This hybrid ML framework consists of random forest (RF), gradient boosting (GB), and artificial neural network of PyTorch library (ANNPT). The final predicted HSAF is effectively reproduced from one training algorithm to another, following the newly introduced implementation for the concept of multi-stage ensemble framework. We used two types of training datasets. Regional dataset consists of EHVSRs at 75 earthquake stations in Egypt. Local dataset consists of MHVSRs at 103 microtremor measurement sites in the Nile Delta basin, northern Egypt. Thus, we can achieve two hybrid ML models, namely regional and local models. Additionally, we combined both regional and local datasets to achieve a combined model. We evaluated the predictive performance of these three models using target or unseen datasets, which comprise MHVSRs at 70 microtremor measurement sites, to examine the potential of applying these models to any given site. Another improvement to the hybrid RF→GB→ANNPT framework is proposed by introducing a constrained condition for the predicted HSAF to be ≤ three times the MHVSR. Thus, we can achieve the best predictive performance using the combined training dataset. The quantitative indicators of confidence, root mean square error, and mean absolute error are 96.5%, 1.01, and 0.78, respectively.
Landslide identification using automated techniques helps researchers improve the accuracy of state-of-the-art landslide prediction models. In recent years, convolutional neural networks (CNNs) have seen considerable success in analyzing remote-sensing images. Its shortcomings in long-range modeling, however, are unfavorable for super-resolution images with speckle noise and shadows and lead to a reduction in the segmentation accuracy of the landslide region. The transformer can gather enough global data, but it struggles to get enough local information and needs to be trained on a huge amount of data in advance. This paper uses a Hybrid CNN-Transformer network to boost the landslide region segmentation in super-resolution remote sensing images. Instead of providing images directly, as reported in prior studies, we employ the feature map generated by the visual saliency as the input to this network. Extensive tests on three publicly available landslide datasets show that the proposed model performs better on landslide region segmentation than existing remote sensing image segmentation methods and the most recent semantic segmentation approaches.
Globally, advanced forest segmentation methods are essential for optimal environmental monitoring, managing resources and ecological studies. As, these techniques uses high-resolution satellite and aerial images for accurate description of forest boundaries and measures the ecological condition. Conversely, traditional segmentation techniques frequently come cross substantial encounters, which includes managing dense vegetation, variation in light conditions and occurrence of shadows that can affects the accuracy of feature detection and segmentation. For these issues, an Iterative Bilateral proposed U-Net with Spatial Pyramid Pooling Layer model is used, which improves U-Net structure by the integration of iterative bilateral filtering. Furthermore, the proposed model improves the segments the forest region by preserving edges with the effective noise reduction and hence it increases the clearness of the segmented output. However, the high-resolution aerial images from various forest areas exactly interpreted and it is used by the proposed model for training and validation. Hence, the model has obtained an accuracy of 85.26% and demonstrate substantial enhancement in segmentation when compared with traditional methods. Besides, it has achieved higher Intersection over Union (IoU) scores that indicates higher overlap among predicted and ground truths. In addition to that the proposed U-Net model excelled in detecting small and more intricate features in the forest areas that are ignored by the conventional methods. Besides, it has an ability to handle high reliable in edge detection when there is an occurrence of noise effectually that is difficult in densely vegetated regions. Thus, the proposed model is computationally effective, generating it to be applicable for real-time forest monitoring and management.
The initial arrival time of seismic waves serves as crucial data in seismic information, efficiently depicting underground velocity structures. The Eikonal equation, a nonlinear partial differential equation utilized for modeling seismic first-arrival travel-times, illustrates how seismic waves propagate in heterogeneous 3D velocity environments. Recent research has shown that Physics-Informed Neural Networks (PINN) have successfully solved specific partial differential equations. However, when applied to the Eikonal equation, PINN exhibit error accumulation that increases with propagation distance. This limitation affects their ability to accurately model seismic wave travel times over longer distances, highlighting the need for improved methodologies to enhance the precision of seismic data interpretation. To address this limitation and the sensitivity to loss function weighting in PINN, we propose a novel multi-agent reinforcement learning (MARL) enhanced PINN approach for solving the Eikonal equation. Our method introduces two agents that dynamically adjust the loss weights (specifically, the weights for the partial differential equation loss and the boundary condition loss) during training. Each agent employs a Deep Q-Network (DQN) algorithm to update its policy, aiming to minimize the overall loss. We rigorously tested our approach on 3D velocity models to validate its robustness and generalizability. Comparative evaluations on conventional geological velocity models demonstrate that our method achieves higher accuracy than standard PINN approaches while effectively mitigating the error accumulation observed with increasing distance from the source.
This study aims to evaluate various approaches for forecasting maize prices by applying individual and ensemble models, both with and without time series decomposition techniques. In the first phase, four decomposition methods were applied: Empirical Wavelet Transform (EWT), Variational Mode Decomposition (VMD), Singular Spectrum Analysis (SSA), and EMD. Subsequently, the signal-to-noise ratio (SNR) was calculated for each technique, and SSA and EMD were selected as the most efficient methods for extracting relevant components and reducing noise. Based on the two selected decomposition techniques, EMD yielded the most favorable forecasting performance. Under this approach, the LightGBM model achieved an RMSE of 3.27, an MAE of 2.08, and a MAPE of 0.50%, while XGBoost obtained an RMSE of 3.70, an MAE of 2.41, and a MAPE of 0.58%. Under SSA decomposition, both tree-based models maintained competitive predictive performance, although with slightly higher forecasting errors. In contrast, the Fully Connected Network (FCN) and Recurrent Neural Network (RNN) models exhibited lower predictive accuracy across both decomposition methods. In a subsequent stage, the PSO–CS optimization strategy was employed to combine individual forecasts, with the XGBoost + LightGBM ensemble under EMD achieving the lowest prediction errors (RMSE = 3.20, MAE = 2.03, and MAPE = 0.48%). Overall, tree-based models consistently outperformed neural network architectures, demonstrating strong predictive capability and robustness, even when trained directly on the original time series without decomposition (MAPE = 0.61%).
Soil texture classes (STCs) can be digitally mapped by first estimating particle-size fractions (PSFs). This study focuses on developing an extreme gradient boosting (XGBoost) model to estimate the spatial distribution of PSFs for a soil depth of 0–50 cm. A set of 30 predictors categorized into main terrain-related attributes, RS data derived from Landsat 9 OLI, climate, and decoded surficial soil color from hue-value-chroma to red-green-blue was used (hereafter referred to as the total dataset). Principal component analysis was performed to minimize the size of the predictors (hereafter referred to as the minimum dataset). In addition to the spatial distribution map of PSFs, the 95% confidence intervals were mapped. Subsequently, basic and sub-categorized STCs were mapped for the entire study area. The spatial intersection of soil texture classes with three landscape features was examined: land use/land cover (LULC), terrain morphological units (TMUs), and slope aspects (north-ward, south-ward, east-ward, and west-ward). Based on 5-fold cross-validation, sand and clay were estimated more accurately using the minimum set of predictor proxies. Allocating original STCs as a reference to four classification systems revealed that the Canadian and USDA systems have the highest agreement (kappa = 0.51 and 0.49, OA = 64% and 63%), followed by the UK and ISSS systems. Nevertheless, the USDA system, widely adopted internationally, was used for mapping. According to the results, the XGBoost model produced more accurate estimates for the sandy clay and sandy clay loam texture classes. Additionally, a higher clay content was observed on the northern slopes compared with the southern slope; grassland soils indicated a 2.5-fold greater area of fine texture and a 2.45-fold lower area of coarse texture than cropland soils; and a 2.27-fold high level of fine texture was found in the north-ward (NW: 0°- 45° and 315°-360°) compared to the south-ward (SW: 135°-225°) over the 0–50 cm depth. It was ultimately concluded that including soil color as a low-cost proxy predictor not only improves PSF estimation accuracy but also, through feature selection, enhances the accuracy of soil texture mapping across the study area located in northwestern Iran.
Geochemistry plays a central role in addressing planetary-scale challenges, yet its scientific practice remains constrained by fragmented, artisanal workflows that limit reproducibility, throughput, and the generation of AI-ready datasets. Although advances in automation, robotics, and artificial intelligence have transformed fields such as genomics, materials science, and chemical synthesis, their adoption in geochemistry has been uneven and largely confined to isolated “islands of automation”. This Perspective introduces the iGeochem Cloud, a comprehensive cyber–physical architecture designed to unify these emerging capabilities into an integrated framework for what we term Intelligent Automated Geochemical Laboratories (IAGLs). The system is centered on a provenance-rich Digital Core that captures complete experimental context, an AI/ML Orchestration Layer that scales expert decision-making, and a FAIR Data Fabric that enables interoperability across laboratories, sensor networks, and institutional boundaries. Together, these components convert traditionally manual workflows into reproducible, automated, and data-intensive research pipelines. We argue that this infrastructure not only addresses current methodological limitations but also establishes the conditions required to develop a domain-specific Geochemical Large Language Model (Geo-LLM). By generating high-fidelity, globally harmonized datasets, the iGeochem Cloud provides a foundation for AI-driven hypothesis generation, experimental design, and interpretation, heralding a paradigm shift toward autonomous and collaborative geochemical discovery. This Perspective outlines the scientific rationale, technical architecture, and implementation roadmap needed to realize this vision and calls for collective community action to build an equitable, sustainable future for intelligent geochemistry.
This study investigates the bearing-capacity factor (Nγ) of rigid strip footings placed on dense sand slopes by integrating finite element limit analysis (FELA) with machine learning and a symbolic regression approach. Two-dimensional FELA was performed under plane-strain conditions using Bolton model, and the numerical results were compared with established results from previous studies to verify the consistency of the predicted trends. Six input parameters were examined: footing width (B), particle crushing strength (Q), relative density (DR), critical-state friction angle (ϕcv), slope angle (β), and slope-height ratio (H/B). The FELA-generated database was used to develop three predictive models: XGBoost, Random Forest (RF), and Evolutionary Polynomial Regression with Multi-Objective Genetic Algorithm (EPR Moga-XL). The EPR Moga-XL model provided explicit and interpretable equations for preliminary design, RF served as an ensemble-based benchmark, and XGBoost achieved the strong overall predictive accuracy, with testing R2 values of 0.994, 0.943, and 0.953 for β = 15°, 30°, and 45°, respectively. The parametric and SHAP analyses showed that Q, DR, B, and ϕcv, are the most influential factors controlling Nγ, while the influence of H/B is strongly dependent on slope angle and becomes more pronounced under steeper slope conditions. Larger B reduces Nγ through stress-level-dependent suppression of dilatancy, while steeper slopes restrict passive-zone development and promote localized failure along the slope face. A supplementary FELA assessment showed that the footing-to-slope crest distance ratio (L/B) strongly affects Nγ near the slope crest, with the response approaching level-ground behavior when L/B = 6. Overall, the proposed FELA–ML framework provides a physics-informed and computationally efficient tool for predicting the bearing capacity of rigid strip footings on dense sand slopes within the adopted applicability boundaries.
Drought is among the most destructive hydroclimatic hazards, particularly in arid and semi-arid regions, where water scarcity directly threatens agricultural production and socioeconomic stability. Reliable seasonal drought prediction is therefore essential for effective early warning and water resources management. This study proposes a sequential hybrid prediction framework, termed Prophet–LSTM–BPNN, that integrates the Prophet model, a long short-term memory (LSTM) network, and a backpropagation neural network (BPNN) for seasonal drought prediction in Iran. The framework is applied separately to monthly basin-averaged values of the three-month standardized precipitation evapotranspiration Index (SPEI-3) for 30 major hydrological basins across Iran during 1990–2021, relying exclusively on univariate time-series data to ensure applicability in data-limited contexts. The modeling strategy hierarchically decomposes drought dynamics: Prophet captures dominant trends and seasonality, LSTM models the remaining nonlinear temporal structures, and the BPNN combines these intermediate outputs. Prediction skill is assessed using a five-fold rolling-origin expanding-window validation scheme, in which each model forecasts 12 monthly SPEI-3 values for five predefined target years: 2011, 2013, 2015, 2017, and 2019. Model performance is evaluated using the Nash-Sutcliffe efficiency (NSE), coefficient of determination (R2), and root mean square error (RMSE). Performance is benchmarked against standalone and hybrid baselines, including Prophet, LSTM, Prophet–LSTM–Add, and Prophet–LSTM–Lin models. The results demonstrate that Prophet–LSTM–BPNN outperforms all baseline models, achieving a mean RMSE of 0.329 and a mean NSE of 0.833 across the 30 basins. Notably, it surpasses the linearly weighted hybrid model, Prophet–LSTM–Lin (RMSE = 0.359, NSE = 0.800), confirming that nonlinear fusion is essential for capturing complex interactions between trend and residual components. This study provides methodological guidance for seasonal drought prediction in data-limited and water-scarce regions and highlights the suitability of the Prophet–LSTM–BPNN framework as a predictive tool for basin-scale drought early warning.
Climate change has continued to be a pressing concern, particularly for climate-vulnerable regions around the world, including Sub-Saharan Africa. While climate change is occurring at a rapid rate, its impact increases in severity. This article examines the biophysical assessment of climate change severity in the Upper East Region (UER) of Ghana. The study’s methodology combines satellite remote sensing data, field surveys, participatory mapping, and policy analysis. A year-long mean, Normalised Differential Vegetation Index (NDVI), rainfall pattern, evapotranspiration (ETa), and Land Surface Temperature (LST) were generated for four different years from Google Earth Engine. A survey involving 200 respondents was also conducted to complement the satellite-based imagery. The findings of the study showed a decline in the mean annual NDVI from 0.40 in 1990 to 0.35 in 2024, thus indicating a decline of 14% in vegetation health. The NDVI decline reveals a transition in vegetation from Sudan to Sahel savannah, suggesting an increasing climate change impact in the region. Rainfall also indicated a downward trend, thus reducing from 1216mm/yr in 1990 to 1089mm/yr in 2024. Also, we discovered an upward trend in LST, with mean maximum LST increasing from 39 0C in 2013 to 45 0C in 2024 and mean minimum LST increasing from 310C to 330C. In the same trend, ETa results suggest increasing evapotranspiration from a mean maximum of 113mm/yr in 2000 to 178mm/yr in 2024. While satellite-based data has revealed worsening climatic conditions in the UER, smallholder farmers acknowledge this and lament that its impact threatens their livelihoods and food security in the region. To address the situation, proactive climate action through collective effort, institutional commitment, and long-term localised adaptation strategies has been recommended.
Post-stack seismic inversion plays a crucial role in the quantitative interpretation of reservoirs. Intelligent inversion approaches employ neural networks to characterize the relationship between observed seismic records and acoustic impedance, thereby avoiding prior assumptions regarding the inverted impedance. However, the scarcity of logging data often hinders the adaptability of inversion networks to field data. Transfer learning offers a feasible solution to this issue by pre-training the network with synthetic data and fine-tuning it with field data. Nevertheless, this approach remains constrained by the need for network to simultaneously adapt to the characteristics of subsurface impedance and the seismic records. To mitigate the dependence on sample availability, we propose a source-independent transfer learning framework for post-stack seismic inversion. Based on the convolutional seismic forward model, we derive a novel source-independent forward operator by incorporating well-log reflectivity sequences. This unified forward model ensures consistency between the synthetic pre-training and field fine-tuning stages, effectively eliminating the influence of seismic wavelet discrepancies on inversion results. Tests on both synthetic and field data demonstrate the effectiveness of the proposed method, confirming its improved stability and accuracy in practical applications.
In Indonesia, wildfires have become an annual disaster that results in significant losses across various aspects of life, including ecological, social, and economic conditions. To minimize these losses, accurate wildfire predictions are urgently needed for prevention, early detection, and wildfire management decision support. This study employs an ensemble learning approach to develop a prediction model for wildfire occurrences and create a susceptibility map of fire-prone areas on a national scale in Indonesia. The proposed model is Stacking Ensemble Learning (SEL), which integrates K-Nearest Neighbor (KNN), Adaptive Boosting (AdaBoost), and Extreme Gradient Boosting (XGBoost) as the base learners, with Random Forest (RF) serving as the meta-learner. The final performance results indicate an accuracy 0.985, balanced accuracy 0.97, precision 0.96, average precision 0.896, recall 0.92, F1-Score 0.954, Matthew's Correlation Coefficient (MCC) 0.944, and AUC Score 0.97, with no signs of overfitting and optimal computational efficiency. SHAP Explainable AI analysis is employed to identify the most influential factors, revealing that proximity to residential areas and climate factors are the most significant contributors to wildfire occurrences in Indonesia. The susceptibility mapping results highlight the provinces with the largest and most vulnerable areas: West Kalimantan, South Sumatra, and South Sulawesi. The outcome of this study can assist stakeholders in mitigating wildfires and protecting the environment to achieve sustainable development goals.
